Reinforcement learning agents trained to handle missing data attacks
Admissable: Training Reinforcement Learning Agents against Adversarial Missingness
Machine Learning
Summary
Sometimes, robots or software that learn from their surroundings can lose important information because someone hides or removes parts of what they see or sense. This makes them perform worse or fail. The authors studied this problem, called adversarial missingness, where an attacker blocks observation features intentionally. They created a new training method that helps these agents stay strong even when some information is hidden on purpose. Their approach worked better than regular training on three test tasks with simulated robots.
What this means in practice
- •For robotics developers: Train robots to maintain performance when sensors fail or data is blocked during operation.
- •For autonomous vehicle engineers: Improve self-driving car systems to resist deliberate sensor attacks that hide critical input features.
Authors
Paul Stahlhofen, Luca Hermes, Tim Kochs, Markus Vieth, Barbara Hammer
Abstract
In order to make Reinforcement Learning algorithms applicable in real world scenarios, safety must be ensured even under adverse operating conditions. In this work, we consider the challenge of adversarial feature missingness: a scenario in which an adversary occludes features from the agent's observation in order to reduce performance as much as possible. We formally define adversarial missingness for Reinforcement Learning and compare it to the related concepts of $\ell_\infty$-norm bounded adversarial perturbations and learning with missing data. We develop an adversarial training algorithm and show its effectiveness in increasing robustness against adversarial missingness on three MuJoCo benchmark environments. Compared to a baseline trained with random uniform missingness, our method achieves better robustness on all three tasks.